Professional Documents
Culture Documents
PS4 Shader Language - Richard Stenson
PS4 Shader Language - Richard Stenson
for PlayStation®4
Richard Stenson
Chris Ho
US R&D
1
PS4, PSSL, and Beyond
- Today we will discuss
- The PS4 architecture
- Developing for PS4
- PSSL on PS4
- Beyond PC with PSSL on PS4
- Join the discussion
2
PlayStation®4
- Next Gen PlayStation Console
- Powerful game machine
- Modern Graphics features
- PC based architecture
- Lightning fast Memory
- New networking and
interface features
3
Modern GPU
- DirectX 11.2+/OpenGL 4.4 feature set
- With custom SCE features
- Asynchronous compute architecture
- 800MHz clock, 1.843 TFLOPS
- Greatly expanded shader pipeline compared to PS3™
4
Fast GDDR5 RAM
- 8GB 256 bit GDDR5
- Fully unified address space
- 176 GB/s total bandwidth
- Massively faster than DDR3
- 128 bit at ~40GB/s max bandwidth
5
State of the art CPU
- Modern 64-bit x86 architecture
- 8 cores, 8 HW threads
- Atomics
- Threads
- Fibers
- ULTs (user-level threads)
6
GPU+RAM+CPU = Beyond Fast!
- Plenty of power for a true Next Gen Game Experience
- 8 CPU cores
- High polygon throughput
- High pixel performance
- Efficient branching in
GPU Shaders
7
But what about development?
- PS4 is very approachable for development
- DX11/OpenGL 4.4 level Shader language in PSSL
- Powerful Graphics API
- C++11 CPU Compiler
- All the expected system libraries and utilities
- Networking, Codecs, Controllers, Input and more
8
Familiar PC-like Development Platform
- Full Visual Studio Integration
- Minimal work for good performance
- Built for AAA Games and Indies alike
- Built to enable developers to push the system
- Good is just the start!
- Once you are ready for the deep dive we support you there as well
9
What is PSSL
- PSSL is the PlayStation Shader Language for PS4
- Supports modern graphics development
- Vertex
- Pixel
- Geometry
- Hull
- Domain
- Compute
10
Vertex and Pixel Shaders
- Next generation VS and PS Shaders
- Extended support based on our hardware
- RW_Textures and Atomics in all shaders
11
Geometry Shaders
- Supports special cases GS like
- GS Tessellation
- Instancing
- Cube mapping
- Streamout
12
Hull, and Domain
- Supports HS DS Tessellation
- Parametric surface conversion
- Optimal Geometry generation
13
Compute
- Support modern compute
shaders
- Parallel Multithreaded
execution
- This cross wave and group
synchronization primitives like
barriers and atomics
- Various Local and Global
memory pools for complex
thread interaction
14
What does PSSL look like?
- It follows the PC conventions for shaders
- ANSI C style syntax and coding rules
- Includes the expected:
- Vectors
- Standard libs
- C++ style structs with members
- Supports static and dynamic control flow
15
A simple vertex shader
struct
VS_INPUT
{
float3
Position
:
POSITION;
float3
Normal
:
NORMAL;
float4
Tangent
:
TEXCOORD0;
float2
TextureUV
:
TEXCOORD1;
};
VS_OUTPUT
main(
VS_INPUT
input
)
{
VS_OUTPUT
Output;
Output.Position
=
mul(
float4(input.Position.xyz,1),
m_modelViewProjection
);
…
float3
vN
=
normalize(mul(float4(input.Normal,0),
m_modelView).xyz);
…
return
Output;
}
16
A simple pixel shader
SamplerState
samp0
:
register(s0);
Texture2D
colorMap
:
register(
t0
);
Texture2D
bumpGlossMap
:
register(
t1
);
float4
main(
VS_OUTPUT
In
)
:
S_TARGET_OUTPUT
{
...
float4
diff_col
=
colorMap.Sample(samp0,
In.TextureUV.xy);
float3
spec_col
=
0.4*normalGloss.w+0.1;
...
return
float4(vLight.xyz,
diff_col.a);
}
17
How PSSL is being developed
- World wide collaborative efforts
- US R&D Shader Technology Group
- PS Vita shader compiler team in ATG
- Graphics driver team in ICE
- GPU hardware teams and SDK managers
- With tight feedback with Sony World Wide Studios
- QA Team
- Thousands of automated tests
18
Let’s see some PSSL shaders in action
- This is real-time PS4 game footage
- All shaders in these demos were built with the PSSL tool
chain
19
The video was removed so this version could be emailed
Video is available upon request or via the ppt version on
devnet
Here https://ps4.scedev.net/support/issue/8907
20
Porting to PSSL from the PC
- Easy initial port target
- Simple conversion of your PC or Xbox 360 Shader
- PS3 Cg conversion is fairly trivial
- Prototyping on the PC much simpler this generation
21
Maintaining PSSL and PC Shaders
- Simpler to maintain code this round
- PC and PS4 are now much closer for shaders
- All of the shader stages and features are available in PSSL
- Often have been extended
- This means you should be up and running very quickly
- The time to “my first tri” will be better
- The time to “my game runs!” will be better
- The time to “my game is fast on PS4” will also be better!
22
Beyond PC with PSSL on PS4
- Extended Buffer Support for all shaders
- Not just Pixel and Compute
- The hardware is capable so we expose that.
- Special Hardware Intrinsics
- Some native ISA instructions are natively supported
- ballot - Good for fine grain Compute control
- sad - For multimedia tasks like Motion Estimation for accelerated image
processing
23
Beyond modern PC shader features
- PS4 GPU has many special features
- Let’s talk about a specific example
24
Example
- New features over previous generation
- New shader stages
- Hull, Domain, Geometry, Compute
- Atomics and RW_Buffers
- Accessible in all stages
- Partially Resident Textures
- What can we do with all of this?
- Why not Sparse Voxel Octree Cone Tracing!
25
Sparse Voxel Octree Cone Tracing
Light
- Global Illumination source
solution proposed by
Crassin et al. in 2011 l
Diffuse
n
- Trace cones through a cones
voxelization of the d
scene to solve for the View
direc3on
n
contribution of direct
and indirect light
sources Images credit Cyril Crassin’s GTC presentation
“Octree-Based Sparse Voxelization for Real-Time
Global Illumination” 26
Sparse Voxel Octree Cone Tracing
- Prepass: voxelize static geometry
- During gameplay:
1. Voxelize dynamic geometry
2. Light volume
3. Build mipmaps
4. Render gbuffers
5. Cone trace scene Quadrilinearly
interpolated
samples
3D
MIP-‐map
pyramid
of
pre-‐filtered
values
Images credit Cyril Crassin’s GTC presentation
“Octree-Based Sparse Voxelization for Real-Time
Global Illumination” 27
Sparse Voxel Octree Cone Tracing
- Could do a full implementation
- (RW_)Texture3D for bricks
- (RW_)RegularBuffer for octree representation
- Geometry shader for thin surface voxelization
- Other useful PSSL features
- Partially Resident Textures?
28
Partially Resident Textures
- Also called “Tiled Resources”
- Hardware Virtual Texturing
- Textures broken up into 64KiB tiles
- Tile texel dimensions dependent on texture
dimensionality and underlying texture format
- Allows for not all the texture to resident in
memory at a time
29
Partially Resident Textures
- Like this, but in hardware!
Physical Representation
Virtual Texture
- For more information, please refer to the Hardware Virtual Texturing
slides presented at SIGGRAPH 2013
30
PSSL and PRT
- Exposed in PSSL as a new Sparse_Texture* type
- All sample-able texture types supported, 1D, 2D, 3D, Cube, Arrays,
etc.
- Sample() modified to take an extra out parameter to
indicate status
- It’s not necessary to use the Sparse_Texture type to utilize
partially resident textures, but Sparse_Texture is necessary
if you want status information!
- Essentially page-fault tolerant GPU memory accesses
31
PSSL Sample Code
Sparse_Texture2D<float4>
sparseTexture;
float4
main(VS_OUT
inv)
:
S_TARGET_OUTPUT0
{
SparseTextureStatus
status;
float4
sampleColor;
//
Try
the
regular
LOD
level
first
sampleColor
=
sparseTexture.Sample(status.code,
sampler1,
inv.tex0);
//
If
'Sample'
fails,
handle
failure
if
(
status.isTexelAbsent()
)
…
32
SparseTextureStatus
struct
SparseTextureStatus
{
uint
code;
bool
isTexelAbsent();
bool
isLodWarning();
uint
getAbsentLod();
//
LOD
of
absent
texel
uint
getWarningLod();
//
LOD
that
caused
the
warning
};
33
PRT Applications
- Megatexturing
- Ptex
- Sparse Voxel Cone Tracing
34
Sparse Voxel Octree Cone Tracing
- Instead of populating an octree, use a partially resident texture!
- Pros:
- PRT tiles do not need to be padded for proper interpolation
- No need to build an octree data structure
- No need to incur the indirection costs of traversing an octree data structure
- Cons:
- PRT tile dimensions not ideal – 64x64x4 for 32-bit 3D textures
- No fast empty space skip from octree traversal
35
Voxelization
- Adaptation of Crassin’s
method detailed in
OpenGL Insights
Hardware
Setup/
VS
Geometry
Shader
Fragment
Shader
Raster.
- Unfortunately no atomic
floats; quantized ints for
accumulation rather than Triangle
Edges
ShiFing
Fragments
Clipping
Voxel
spin lock Dominant
Triangle
A>ributes
Write
to
3D
Axis
Projec;on
Conserva;ve
surface
Computa;on
- Use geometry shader Selec;on
Rasteriza;on
y y
pass x z
37
Tile Allocation
- Map texture initialized to set a reserved “unallocated bit”
- AtomicCmpExchange() in a value to flip on an additional
“unallocated-but-I’m-working-on-it” bit for a single thread
- Consume() a free tile
- Append() consumed tile with remap data
- Write out tile location to map texture
- Write into the tile using pool texture
- After pass completion, read from append buffer on CPU side to
map tiles from the pool to the sparse texture
38
Tile Allocation
39
Tile Allocation
const
uint
unallocated
=
0x80000000,
allocating
=
0xC0000000;
do
{
cur
=
map[tileLoc];
if(cur
==
unallocated)
{
uint
output
=
0xffffffff;
AtomicCmpExchange(map[tileLoc],
unallocated,
allocating,
output);
if(output
==
unallocated)
{
cur
=
g_freeTiles.Consume();
map[tileLoc]
=
cur;
g_remaps.Append(…);
}
}
}
while(cur
&
unallocated);
40
Implementation
- 1024x1024x512 32-bit pool texture
- 16x16x128 tiles, given linear ids (can use shifts/masks to find actual
location)
- 512x512x512 32-bit Sparse Texture to represent the scene
- 8x8x128 map texture for tile allocation
- Consume buffer for grabbing free tiles
- Append buffer for noting allocated tiles for remapping
41
Building Mipmaps
- Compute Kernel that takes an 8x8x8 brick and reduces
it to a 4x4x4 brick
- LDS for accumulating final values
- Allocate tiles for new mips in the same manner as
voxelization
- Pre-map the lowest mips (all that fit into 64KiB)
42
Lighting Voxels
- Currently naively lit
- Spawn Compute kernel for entire 3D texture, iterate
over lights if resident
- Needs optimization
43
Results
44
Results
- Average frame time: 26ms
- 3ms gbuffer, 11ms indirect + specular reflect, 11ms direct
- Memory usage:
- 2GiB Pool Texture, ~315MiB allocated after voxelization, ~56%
resident
- Static geometry voxelization and lighting time:
- 45ms voxelization, 22ms top-mip light, 25ms mip regeneration
- Still much more optimization possible!
45
PSSL is still evolving
- Features in consideration:
- Control of shader resource
layout
- More exotic compute primitives
for GPGPU
- Tightly coupled Graphics and
Compute
- And many more…
46
Join the discussion
- We would like to hear from you!
- Sign up as a PS4 developer, if you’re not already
- http://us.playstation.com/develop/
- There is a link for all territories from this page
- We are looking for solid suggestions with clear benefits
- Specific performance benefit
- Special new/novel feature, etc.
47
Push the boundaries with PSSL
- PS4 is a powerful, but friendly to develop for
- PSSL is one of the keys for developing for PS4
- Our goals with PSSL
- Make better Games
- Push the boundaries on PS4
- And to be efficient in that process
48
Q&A
- Questions?
49
50